feat(api): add support for highres images

api/onnx_web/diffusers/run.py

@@ -6,10 +6,12 @@
 from diffusers import OnnxStableDiffusionImg2ImgPipeline, OnnxStableDiffusionPipeline
 from PIL import Image
 
+from onnx_web.chain.utils import process_tile_order
+
 from ..chain import blend_mask, upscale_outpaint
 from ..chain.base import ChainProgress
 from ..output import save_image, save_params
-from ..params import Border, ImageParams, Size, StageParams, UpscaleParams
+from ..params import Border, ImageParams, Size, StageParams, TileOrder, UpscaleParams
 from ..server import ServerContext
 from ..upscale import run_upscale_correction
 from ..utils import run_gc
@@ -79,6 +81,64 @@ def run_txt2img_pipeline(
         )
 
     for image, output in zip(result.images, outputs):
+        highres_scale = 4
+        highres_strength = 0.5
+
+        if params.highres > 1:
+
+            def highres(tile: Image.Image, dims):
+                highpipe = load_pipeline(
+                    server,
+                    OnnxStableDiffusionImg2ImgPipeline,
+                    params.model,
+                    params.scheduler,
+                    job.get_device(),
+                    params.lpw,
+                    inversions,
+                    loras,
+                )
+                progress = job.get_progress_callback()
+                if params.lpw:
+                    logger.debug("using LPW pipeline for img2img")
+                    rng = torch.manual_seed(params.seed)
+                    result = highpipe.img2img(
+                        tile,
+                        params.prompt,
+                        generator=rng,
+                        guidance_scale=params.cfg,
+                        negative_prompt=params.negative_prompt,
+                        num_images_per_prompt=1,
+                        num_inference_steps=params.steps,
+                        strength=highres_strength,
+                        eta=params.eta,
+                        callback=progress,
+                    )
+                    return result.images[0]
+                else:
+                    rng = np.random.RandomState(params.seed)
+                    result = highpipe(
+                        params.prompt,
+                        tile,
+                        generator=rng,
+                        guidance_scale=params.cfg,
+                        negative_prompt=params.negative_prompt,
+                        num_images_per_prompt=1,
+                        num_inference_steps=params.steps,
+                        strength=highres_strength,
+                        eta=params.eta,
+                        callback=progress,
+                    )
+                    return result.images[0]
+
+            logger.info("running highres fix for %s tiles", highres_scale)
+            image = process_tile_order(
+                TileOrder.grid,
+                image,
+                size.height // highres_scale,
+                highres_scale,
+                [highres],
+            )
+
         image = run_upscale_correction(
             job,
             server,